Repository navigation
spec(hslm): nn/hslm and hslm/forward_pass compile and pass; own_model counts it (Closes #7778) - #10055
Merged
Conversation
…10026) One map for IGLA CODER, HSLM in t27 and the harness, through the T27F topic workflow: statuses earned by topic_engine::earned_status from measurements taken 2026-10-17 on the Railway lab (4 of 17 ML specs compile, 0 of 9 IGLA CODER specs), Hugging Face and the IGLA board. 1 published, 11 verified, 8 drafts, 7 gaps; about 39 percent overall, equal weight per cell. Epic #10024. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#7778) nn/hslm.t27 was BLOCKED on 'constants' (via nn/attention). It is now self-contained, with a real backward pass: every weight, gamma and input gradient is checked against a finite difference or a trit flip. 26/26 pass, 0 vacuous, on the Railway lab. hslm/forward_pass.t27 was BLOCKED on undeclared q_role. It is rewritten on dense ternary hypervectors: bind, bundle, permute, hard attention over positioned context, Hebbian counts and a learned direct role. 14/14 pass, 0 vacuous. own_model.t27: cell 14 (forward pass) now has a passing test and is VERIFIED; 6 of 17 ML specs compile; percents and slides follow. Seals for both specs re-sealed on the lab. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…mpile # Conflicts: # specs/igla/own_model.t27
…7778) own_model now counts the HSLM forward cell as verified: 1 published, 12 verified, 7 drafts, 7 gaps, 42 percent overall, lens means 45/51/30. The board's message, percent, progress, cell arrays and metrics byte count follow; negative controls stay wrong against the new values. Lab: own_model_board 6/6, own_model 8/8, 0 vacuous. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
gHashTag
enabled auto-merge (squash)
October 17, 2026 17:42
This was referenced Oct 17, 2026
gHashTag
added a commit
that referenced
this pull request
Oct 17, 2026
…0071) (#10078) #10055 made specs/nn/hslm.t27 parse with no discarded tokens, so the corpus ratchet on master reports it as an UNEXPECTED PASS (and as a DISCARD WORSENED with no reading). The pin is dropped and max_entries follows the count down, 82 -> 81. The entries match the ledger's own bless command, run on the Railway lab, byte for byte; the bless serializer would also drop the hand-written _why_entries_fell/_why_entries_rose fields, so they are kept and the diff is the one entry plus the cap line. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #7778
What
specs/nn/hslm.t27(BLOCKED: undeclared identifierconstants) andspecs/hslm/forward_pass.t27(BLOCKED: undeclared identifierq_role) now compile and pass. The causes were defects in the specs, not t27c codegen bugs, so only.t27files and seals change.use base::types. The backward pass is real, not placeholders. The FFN, norm and attention gradients are each checked against a finite difference or a trit flip. The old hash put every character in the same mod-9 trit column; a multiplicative hash replaces it.resonatorTrainStepwas deliberately left out (the header explains why).HSLM,nn_HSLM,ForwardPassandhslm_ForwardPass. The old ones had stale spec hashes.Measured (Railway lab, t27c test-report, zig 0.15)
The spec is not the trained model
The spec and the HSLM that trinity-training actually trained (
src/hslm) differ:nn/hslm.t27)Attention matches: 3 heads x 81, pre-RMSNorm, sacred scale, residual. This PR makes the spec compile and pass as written. Aligning it with the trained model is the next row on the own-model board ("Match the spec to the trained HSLM").
🤖 Generated with Claude Code